Co-EvoDT: Design of a Co-evolving Digital Twin Framework for Media Digital Transformation in the Metaverse through Integration of Multi-Agent Reinforcement Learning and LSTM-based Sequential Forecasting

Digital twins in the metaverse have created new opportunities to redefine digital transformation in media; however, the absence of co-evolutionary models that simultaneously optimize infrastructure resource-allocation policies, content production, and load forecasting remains a major barrier to fully realizing these opportunities. To address this gap, we propose the Co-EvoDT model and implement it within a simulated metaverse environment using a design-driven, quantitative research approach. The proposed architecture combines sequential load forecasting via an LSTM predictor (prediction horizon = ۱, sequence length = ۸), multi-agent reinforcement learning trained with the policy-based REINFORCE algorithm to enable co-evolutionary training of content, infrastructure, and user digital twins, and a runtime rolling-update mechanism that appends the LSTM’s normalized output to the state vector used by the control policies. Performance was evaluated using Quality of Experience (QoE), latency, cost, and RMSE of the load predictor. Simulation results show that a policy trained with the combined predictive signal outperforms a random policy by simultaneously improving multiple metrics: a relative increase of ≈۹۳% in QoE, a relative latency reduction of ≈۲۹%, and a cost reduction of ≈۲۷%. Forecast RMSE was reduced by ≈۴۳%, ≈۵۳.۲%, and ≈۸۸% compared to Naive, ARIMA, and Exponential Smoothing baselines, respectively. Reward-curve convergence analysis and parametric sensitivity experiments further corroborate the stability and robustness of the learned policies under variations of key system parameters. The principal innovation of this research is the operational integration of multi-agent reinforcement learning, sequential forecasting, and co-evolutionary population-gradient update mechanisms within digital twins to enable proactive, coordinated resource management in the metaverse. The results indicate that Co-EvoDT can concurrently enhance user experience and macro-level system performance.

کلیدواژه ها:

Media Digital Transformation ، Metaverse ، Digital Twin ، Multi-Agent Reinforcement Learning ، Sequential Forecasting